ee with , tk Φ= + = zA ε fzAy R)(R 2 1 min arg }ˆ,ˆ{ zf fzAy zf +

advertisement
Fast and Motion Robust R2* Reconstruction for functional MRI
V. Olafsson1, J. A. Fessler1, and D. C. Noll2
EECS, The University of Michigan, Ann Arbor, MI, United States, 2BME, The University of Michigan, Ann Arbor, MI, United States
1
Introduction:
Quantitative images for functional MRI (fMRI) are extremely important to further understand the hemodynamic processes of the BOLD
response. In particular R2* maps can be linked to physiological processes such as blood flow, blood volume and metabolic rate of
oxygen, all of which participate in the BOLD response. However, traditional methods to reconstruct R2* maps are either
computationally time consuming or needed excessive amounts of data. A previous approach used a baseline image and a linear
approximation to simplify the problem of estimating both R2* and field map from a single readout to minimizing a quadratic cost function
[1]. This is a fast and accurate R2* reconstruction but assumed the initial magnetization M0 was static for all time frames and thus
ignored motion during an fMRI experiment. Here, we present a fast reconstruction method that also uses repeated linearizations for
speed that reconstructs R2* along with field map and M0, using additional data acquired while waiting for the BOLD contrast.
Theory:
With R2* and off-resonance included, the acquired MR data y that is needed to reconstruct a slice can be modeled as follows:
r
r
r
− t m ⋅ z n − i 2π (k (t m )⋅rn )
(1)
(
)
y = A (z )f + ε, with [A (z )]mn = Φ k (t m ) e
e
,
where A(z) is the system matrix, z is the complex valued rate map with R2* and field map in its real and imaginary parts respectively, f
is the objects initial magnetization M0, Φ(·) comes from the continuous-to-discrete mapping of the MR signal equation [2,3], k(t) is the kspace trajectory and r is the spatial coordinate. We jointly reconstruct f and z using an iterative algorithm for the following regularized
optimization problem:
{fˆ , zˆ } = arg min
f ,z
1
y − A (z )f
2
2
+ R 1 ( f ) + R 2 (z ),
(2)
where R1(f) is a quadratic roughness penalty and R2(z) is a quadratic roughness penalty that separately penalizes R2* and field map [1].
The cost function in (2) is non-convex which can slow iterative algorithms. To accelerate, suppose we add and subtract a previously
acquired reference rate map zref to the exponent with z in (1) and split it into two exponentials, one with zref and the other z-zref. If zref is
close to z, that exponential can be approximated well with its first order Taylor expansion, replacing (2) with:
{fˆ , zˆ } = arg min
f ,z
1
y − A (z ref )f + D(− t )A (z ref )D(f )z ref − D(− t )A (z ref )D(f )z
2
2
+ R 1 (f ) + R 2 (z ),
(3)
where D(·) is a diagonal matrix. Using alternating minimization, this problem reduces to quadratic minimizations for both f and z, which
we solve by using fast iterative algorithms [2,3]. We first estimate f by setting z = zref in (3) and solve, then replace f with that estimate
and estimate z that is then used to update zref, repeating until it converges.
Methods:
We acquired a 2-echo spiral-out data with TE=4.6ms and
30ms, FOV=22cm, FA=55, TR=1.6s, {# of slices}=24. The TEs
were intentionally chosen far apart to minimize correlation
between the estimated z and f images. The task was a visually
cued finger tapping, 20s on 20s off, repeated 5 times. To
estimate the initial zref, we shifted TE of the first time frame by
2ms and estimated the field map from the first echo of the first
and second time frames. We estimated R2* using a log linear
fit on images reconstructed from both echoes of the second
time frame, which was the first time frame of the fMRI
experiment. We reconstructed all time frames of the
experiment using both the proposed method and conventional
T2*-weighted reconstruction of the second echo [3]. All
reconstructed volumes were slice-time and motion corrected
using FSL 4.0 before being analyzed using GLM.
Results and Discussion:
Fig. 1: z-score maps of active areas in the motor (upper) and visual cortex
To solve (3) we ran the alternating minimization 5 times using
(lower) for f (left), R2* (center) and T2*-weighted (right) images
the iterative algorithm proposed in [2]. It took just under a
minute to reconstruct one slice on a 2.6 GHz Intel Xeon. Fig. 1 shows the z-score for voxels above 4.61 for f, R2* and T2*-weighted
images, all overlaid on the reconstructed f for the first time frame. The R2* images seem to boost the overall number of active voxels
compared to the T2*-weighted images. This is probably due to the field map and f being estimated for every time frame, thus correcting
for any changes either due to motion or inflow. Inflow effects are evident from the activation in f. For the inferior slices the estimated f
image has piling up around the sinus and ear canals due to large through plane gradients, which requires further investigation.
Acknowledgement: This work is supported in part by NIH Grants R01NS058576 and R01EB2683
References: [1] Olafsson VT et al, IEEE TMI, 27(9):1177-88, 2008; [2] Sutton BP et al, IEEE TMI, 22(2):178-88, 2003; [3] Fessler JA et
al, IEEE Trans. Sig. Proc., 53(9):3393-402, 2005.
Proc. Intl. Soc. Mag. Reson. Med. 17 (2009)
2831
Download